Halley's Method¶
Halley's method refines a scalar root estimate with a curvature-corrected rational step that converges cubically near a suitable simple root.
Core Idea¶
Halley's method refines a guess \(x_n\) for a simple zero of a nonlinear scalar function \(f\) by using the residual, slope and curvature at that guess:
This is algebraically NIST DLMF's Halley rule. Where the root is simple, the needed derivatives are regular and the iterates begin sufficiently close, the error is ordinarily cubic in the preceding error. “Cubic” is a local convergence order, not an assurance that any initial guess converges or that the denominator can never vanish.[1][2]
The frozen candidate “Bailey's method (root finding)” resolves to this Wikipedia page; its candidate ID remains attached as redirect provenance. That resolution does not establish a separate Bailey algorithm or require using Bailey as a mathematical alias.
Structural Signature¶
Sig role-phrases:
- Nonlinear residual: \(f(x_n)\) states how far the current guess is from satisfying \(f(x)=0\); absent a root equation there is no target.
- Slope and curvature: \(f'\) and \(f''\) set the local step. Removing \(f''\) reduces the formula to Newton's update wherever \(f'\) is defined.
- Iterative correction: the rational fraction is subtracted repeatedly, with values reevaluated at each new iterate. A one-time curvature estimate is not the iteration.
- Local regularity boundary: a simple target root, adequate smoothness, suitable starting region and nonzero update denominator support the characteristic behavior. Those qualifications are part of interpreting the method, not merely implementation trivia.[1][2]
Condensed: residual + slope + curvature → rational root correction → rapid local refinement, conditional on regularity.
What It Is Not¶
- Not Newton's method. Newton uses \(f/f'\); Halley modifies that step with \(f''\).
- Not any second-derivative root solver. Another curvature formula might have a different convergence basin and step.
- Not a globally convergent method. Cubic local order gives no safe-start rule for an arbitrary nonlinear function.
- Not a guaranteed three correct digits per iteration. Order concerns asymptotic error powers, not a fixed decimal gain.
- Not automatically preferable to Newton. Computing a second derivative may cost more than the iterations saved; a small denominator may require safeguarding.
Scope of Application¶
NIST DLMF explicitly presents Halley's rule among methods for zeros of analytic functions. In its Bessel computation chapter it lists Newton or Halley for high-accuracy zeros of functions in that chapter. In its Airy computation chapter it similarly lists Newton or Halley after appropriate asymptotic initial approximations. These are two distinct special-function families using the same update, not evidence that a particular implementation or starting guess works for all functions.[1][3][4]
In software, SciPy's scalar newton interface chooses Halley's method when callers provide the second derivative as well as the first. That is an implementation-selection fact, not a proof that providing \(f''\) always improves running time.[5]
Clarity¶
Newton's method takes a tangent-line estimate. Halley's step uses curvature to correct the tangent prediction. At a suitable nearby simple root, the added derivative information cancels a lower-order error term, explaining cubic rather than quadratic local convergence. The denominator simultaneously reveals a failure mode: if \(2(f')^2-ff''\) approaches zero at a proposed iterate, the rational correction can be huge or undefined. A solver should inspect its step and residual instead of invoking “higher order” as a safety claim.[1][2]
Manages Complexity¶
Finding a nonlinear zero is reduced to repeatable evaluations of \(f\), \(f'\) and \(f''\), rather than constructing a fresh symbolic solution. This makes the method useful when derivatives are accessible and high precision matters. The computation still needs a start, stopping rule and failure handling. For a function with expensive curvature evaluation, a cheaper method might reach a usable answer sooner despite requiring more iterations.
Abstract Reasoning¶
For a constructed example \(f(x)=x^2-2\), at \(x_0=1.5\) the values are \(f=0.25\), \(f'=3\), \(f''=2\). The Halley correction is \(2(0.25)(3)/(2(3)^2-(0.25)(2))=1.5/17.5\), yielding \(x_1=1.414285714\ldots\), close to \(\sqrt2\). This calculation demonstrates the rule but not a theorem about starts outside a root neighborhood. At a simple root \(\alpha\), \(f'(\alpha)\ne0\); as \(x\) approaches \(\alpha\), the denominator approaches \(2[f'(\alpha)]^2\), so the local step remains defined. A multiple root violates that simple-root argument and can have different order.[1][2]
Knowledge Transfer¶
The formula transfers literally from zeros of Bessel-family functions to zeros of Airy-family functions: the function and derivatives change, the update roles do not. Their initial-guess strategies and evaluation costs do change. The abstract idea “use more local derivatives to accelerate an iteration” travels farther, but it is broader than the named Halley step and does not license classifying every high-order method as Halley's method.
Examples¶
A calculated square-root-of-two step¶
Take the concrete equation \(f(x)=x^2-2=0\) with \(x_0=3/2\). Then \(f(x_0)=1/4\), \(f'(x_0)=3\) and \(f''(x_0)=2\). The Halley numerator is \(2(1/4)(3)=3/2\) and its denominator is \(2(3)^2-(1/4)(2)=35/2\), so the correction is \(3/35\). One actual iteration gives \(x_1=3/2-3/35=99/70\approx1.414285714\), compared with \(\sqrt2\approx1.414213562\). The denominator is nonzero at this start; nothing about this one successful step establishes a global basin of convergence.[1]
Mapped back: \(x^2-2\) is the nonlinear residual, the values 3 and 2 are slope and curvature, the exact \(3/35\) fraction is the repeated-rule correction for this first step, and the nonzero \(35/2\) denominator plus a suitable nearby start are the local-regularity conditions.
The documented SciPy cubic-root step¶
SciPy's own Halley example supplies \(f(x)=x^3-1\), \(x_0=1.5\), fprime=lambda x: 3*x**2 and fprime2=lambda x: 6*x, and reports convergence to the root 1.[5] Its first step can be checked from NIST's formula: at \(3/2\), \(f=19/8\), \(f'=27/4\) and \(f''=9\). The numerator is \(513/16\); the denominator is \(279/4\); hence the correction is \(57/124\) and \(x_1=129/124\approx1.040322581\). This is a documented execution target with a derived first iterate, not merely a function family for which Halley might be used.[1][5]
Mapped back: the cubic polynomial is the residual; the two callbacks supply slope and curvature; the \(57/124\) correction executes the Halley rule; the nonzero \(279/4\) denominator and nearby simple root bound the local claim.
Structural Tensions¶
Local rate versus per-step cost. The curvature term can lower error quickly near a root, but evaluating \(f''\) may be expensive or numerically noisy. Counting iterations without counting derivative cost gives a misleading comparison with Newton. Diagnostic: is curvature cheap and stable enough at the requested accuracy to offset the extra work?[1][5]
Aggressive correction versus safeguarded reliability. A pure Halley step can exploit cubic local convergence; a distant guess or small denominator can instead cause a large excursion. A bracket or fallback step may sacrifice local speed to protect progress. Diagnostic: is the iterate inside a trustworthy root neighborhood, and is the correction denominator comfortably nonzero?[1][2]
Structural–Framed Character¶
Halley's method is strongly structural: its residual, derivative roles and correction formula fix an algorithm independent of who runs it. Evaluative weight enters when deciding if extra derivative work is worthwhile, not in the update itself. Human numerical practice supplies starting approximations, accuracy goals and safeguards; no institution constitutes the formula, though NIST and library documentation standardize its presentation. Vocabulary travels literally from Bessel to Airy zeros where the exact step is used. Importing “Halley” to any fast solver merely because it is third order would confuse analogy with recognition. Its character: a formal, domain-specific numerical iteration with a high local rate and an essential start/denominator boundary.
Structural Core vs. Domain Accent¶
The skeletal relation is use successive local derivative information to correct a root estimate. The live domain-specific Iterative Method is a verified strict numerical-method genus; no root-iteration prime is asserted as a parent, and a generalized derivative-based refinement prime would require separate admission. The domain-bound mechanism is the exact rational combination of \(f\), \(f'\) and \(f''\) for scalar nonlinear zeros. The named entry fails the prime bar because changing those ingredients yields Newton, another Householder-type method, or an unrelated optimizer rather than the same Halley algorithm. Broad “iterative improvement” is portable only as a future-prime question, not a fabricated edge.
Instantiates / Related Primes¶
This entry is a kind of Iterative method.
Live Iterative Method is the strict parent: Halley's recurrence is a particular method of successive numerical approximation, although a failed run has no convergence guarantee. Live Gauss–Newton Algorithm is not a parent: it solves a least-squares problem by a different approximation and update. No root-finding prime parent is asserted.
Relationships to Other Abstractions¶
Current abstraction Halley's Method Domain-specific
Parents (1) — more general patterns this builds on
-
Halley's Method is a kind of Iterative method Domain-specific
Halley's root update is a specific iterative numerical method with a curvature-corrected recurrence.Every Halley method repeatedly updates a scalar root approximation from its prior value using the residual, first derivative and second derivative, under a stated start, denominator and stopping or failure regime. Iterative Method is the live numerical-method genus for such successive approximations. Newton's method and iterative linear solvers show the parent can exist without Halley's rational correction. The edge concerns method identity and conditional local convergence, not a guarantee that every run succeeds.
Hierarchy path (1) — routes to 1 parentless root
- Halley's Method → Iterative method → Iteration
Neighborhood in Abstraction Space¶
Halley's Method sits in a sparse region of the domain-specific corpus (88th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Ridders' Method — 0.83
- Gauss–Newton Algorithm — 0.82
- Euler Method — 0.80
- Bisection Method — 0.80
- Limited-Memory BFGS — 0.80
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Newton iteration: first derivative only. Generic Householder iteration: family-level comparison, not an asserted catalog identity. A cubic polynomial solver: solves a different problem. Bailey's method (root finding): frozen redirect route to the resolved Wikipedia article, retained as provenance rather than promoted to a sourced synonym.
References¶
[1] NIST DLMF, §3.8(v), Eq. 3.8.12, Halley's rule. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i
[2] NIST DLMF, §3.8(i), simple zeros and local convergence order. registry ↩a ↩b ↩c ↩d ↩e
[3] NIST DLMF, §10.74(vi), Bessel-family zero computations. registry ↩
[4] NIST DLMF, §9.17(iv), Airy-function zeros and initial approximations. registry ↩
[5] SciPy, optimize.newton documentation, selection of Halley when fprime2 is supplied. registry ↩a ↩b ↩c ↩d